Electric Power and Energy Sates
Vol. 1b, Note Feb. 1972, ppg 17
An optimization-based method for
unit commitment
X Guan, P B Luh and H Yan
Department of Electrical and Systems Engineering,
University of Connecticut, Storrs, CT 06263-3167.
USA
JA Ama
Northeast Utilities Service Company, Berlin, CT
06037-1616, USA,
‘An optimization-based method for unit commitment using
the Lagrangian relaxation technique is presented. The
salient features of this method includes nondiscretization
of genération levels, a systematic method to handle ramp
rate constraints, and a good initialization procedure. By
sing Lagrange multipliers to relax system-wide demand
‘and reserve requirements and ramp rate constraints, the
problem is decomposed into the scheduling of individual
tunits. The optimal generation level of a unit at each hour
can be easily calculated since there are no system dynamics,
and the cost function is stage-wise additive and piecewise
linear with only a few corner points. A relaxed subproblem
can therefore be efficiently solved by using the dynamic
Programming technique without diseretizing generation
levels. A subgradient algorithm with adaptive step sizing
isused to update Lagrange multipliers. An effective method
based on priority-list commitment and dispatch is adopted
10 initialize these multipliers, and a heuristic approach is
developed to generate a good feasible schedule based on
the dual solution. Numerical results based on data sets
from Northeast Utilities show that this algorithm is
‘ffcient, and near-optimal solutions are obtained.
Keywords: unit commitment, power system scheduling,
‘mathematical programming, Lagrangian relaxation
|. Introduction
Unit commitment of a thermal power system is used to
determine when to start up and/or shut down thermal
units, and how to dispatch the committed units to mect,
system-wide demand and reserve requirements over a
period of up to one week. Each unit may have minimum
up and down times, ramp rate and other constraints.
This class of mixed integer programming problems has
been an active research subject for several decades
Received 1 August 1991; revised & November 1991; accepted 6
February 1992
because of potential cost savings. According to a recent
analysis, a 1% reduction in operating costs can result in
savings of 10 to 30 million USS per year for an electrical
utility with 10000 MW of installed capacity'. However,
consistently generating optimal schedules has proved to
be very difficult, because the problem belongs to the class
of NP-hard combinatorial problems, and is considered
to be extremely difficult to solve for systems of practical
size (€-g. 100 units).
Recently, impressive results have been obtained by
using the Lagrangian relaxation approach for obtaining
near optimal solutions-®. The basic idea is to relax
system-wide demand and reserve requirements by using
Lagrange multipliers. The problem can then be
decomposed into individual unit commitment sub-
problems, which are much easier to solve. The high-level
problem is to optimize Lagrange multipliers, and can be
solved efficiently by using continuous variable optimiza~
tion techniques. The disadvantage of this method is that
the dual solution is generally infeasible, ic. the once
relaxed system-wide constraints are not satisfied. Some
techniques, usually heuristics, are needed to modify the
dual solution to obtain a good feasible schedule.
Nevertheless, since the value of the dual function is a
lower bound on the optimal cost, the quality of the
feasible solution can be quantitatively measured.
‘A. method for unit commitment based on the
Lagrangian relaxation technique is presented here. The
salient features of this method include nondiscretization
of generation levels, a systematic method to handle ramp
rate constraints, and a good initialization procedure. For
a relaxed subproblem without ramp rate constraint, the
optimal generation level of each hour can easily be
determined since there are no system dynamics, and the
cost function is stage-wise additive and piecewise linear
with only a few comer points. This subproblem is solved
by first constructing a state transition diagram where the
optimal generation levels of all up states are computed
without discretizing generation levels. Dynamic. pro-
‘gramming technique is then applied with only a few well
structured states. This eliminates the difficult trade-offbetween computational requirements and accuracy as
needed by most approaches that discretize generation
levels.
The ramp rate constraint of a unit couples the
generation levels of two consecutive hours. If the
‘generation levels are discretized, this constraint can be
‘bandied by using the standard dynamic programming.
technique®®, The computational requirements, however,
would increase significantly (possibly by an order of
‘magnitude) as compared to the case without ramp rate
constraints?. If the generation levels are not discretized,
the constraint is very difficult to deal with. A
straightforward application of the dynamic programming
technique may lead to suboptimal results (see Sections
ML2 and 111.7). This constraint is handled by using an
‘ad hoc approach in Reference 8, and is not considered in
Reference 9. In Reference 10, a subproblem is solved by
dynamic programming with an extended state space for
ramp-down logic, where limits of generation levels at
each hour are established. The optimality and feasibility
of the solution, however, are not guaranteed. Rather,
ramp constraints are satisfied in a sophisticated econom
dispatch process by using a ‘look-ahead’ heurist
method.
In our paper, ramp rate constraints are relaxed by
introducing an additional set of multipliers for a unit
with the constraints. The subproblem is then solved as
if there were no ramp rate constraint. An intermediate
level is introduced to update this set of multipliers, and
a three-level framework is formed as shown in Figure 1
Note that intermediate subproblems are only needed for
units with ramp rate constraints. An efficient subgradient
algorithm’! is modified to update Lagrange multipliers
associated with system demand, reserve and ramp rate
constraints. An effective method is developed to initialize
these, multipliers based on the priority-list_ commit-
‘ment®#? and dispatch. Since a dual solution is generally
infeasible, a heuristic method is also developed to
construct a feasible schedule based on the dual result.
The purpose of the research is to generate hourly
schedules for Northeast Utilities Service Company (NU)
which has about 70 thermal units. The maximum time
horizon is 240 h (10 days). Numerical results based on
NU data show that this method is efficient, and
Undote mtipiers associated
with demond nd reserve requirements
ft tf
‘Units with rom rote ‘Unite without romp rote
constraints ceonatroints
Update mutters associated Solve ldividuat
th romp rote constants tow level subproblens
ia
‘Solve indivi
tow teal subproblens
Figure 1. Three-level
relaxation
framework of Lagrangian
near-optimal solutions are obtained. The algorithm
developed has been embedded in the daily scheduling
package of NU and used by NU engineers, mathematical
problem formulation is given in Section TI.
Ul, Problem formulation
Consider a thermal power system with I units. It is
required to determine the startup, shutdown, and
generation levels of all units over a specified time period
TT. The objective is to minimize the total cost subject to
system demand and spinning reserve requirements, and
other individual unit constraints. The time unit is one
hour and the planning horizon may vary from one day
to ten days. To formulate the problem mathematically,
the following notation is introduced,
Cpe) fuel cost of unit i for generating power
pi(t) at time t, a piecewise linear
function of p(t), in USS
i index of units, { a
1 number of units
Plt) system demand at time t, in MW
pale) power generated by unit / at time ¢, in
MW
Be) maximum generation level of unit i at
time ¢, in MW.
lt) minimum generation level of unit i at
time ¢, in MW
Pe) system spinning reserve requirement at
time t, in MW *
i ‘maximum spinning reserve contribution
of unit i, in MW
spinning reserve contribution of unit
Gale), PAD)
iat time t, 7(-)=0 if unit is
down (x,(1)<0) and 1,(-)
min{pi(t)— p(t). A} if unit is up
((1) > 0), in MW
amp rate of unit i, in MW/h
startup cost of unit i, in USS
R
Sie (t), u(t)
s
4 cold start-up cost of unit i, in USS
s hot start-up cost of unit i, in USS
t time index, ¢= 1,..., 77
T time horizon of commitment, in hours
(0) discrete decision variable of unit i at
time ¢ for up (1) or down (~1) of the
unit at time ¢+ 1
state of unit iat time ¢, denoting number
of hours that unit i has been on (positive
values) of off (negative values)
maximum allowable change in genera-
tion between two consecutive hours,
4, = 1-R, in MW
minimum up-time of unit i, in hours
minimum down-time of unit,
cold start-up time of unit , in hours
‘The problem is then formulated as the following mixed
integer programming problem
ass
@)
min J, with F= YY CGA) + SC), wl]
*
)
subject to system-wide constraints which include thefollowing
System demand
Y pit) = Pale) )
Spinning reserve
4 mC), PE) > Pet) GB)
Individual unit constraints include
State transition
att = x(t) tut) ifx(e)-m()>0 (4)
x(t I= u(t) ix (et) u(t) <0 ()
i.e. the number of hours being up or down accumulates
‘no start-up or shut-down occurs, otherwise the number
of hours being up or down equals 1
Capacity
BA) S Put) < p(t) f(t) >0 G)
pt) =0 ifx(t) <0 a
‘Some units may also have one or more of the
following constraints.
Ramp rate
C(t +1)— A) < p(t) < Cott + 1) + A)
if x(t) > 1 and x(t+1)>1 (8)
Minimum up/down time
u(t il 0
or ifx(t)>Oandxj(t+1)<0 (11)
Must-run or must-not-run
x(t) >0 for ty 0, t=1,2....,7 (as)
‘The above derivation presents a decomposition
framework for solving the unit commitment problem.
There are several steps to obtaining a near optimal
solution : solving subproblems, solving the dual problem,
and constructing a feasible solution, which are considered
below.
WL2 Solving individual unit subproblems
The solution methodology for a subproblem without
ramp rate constraints is presented first. For the cost
function in equation (17) with 2 and given, define the
non-start-up cost as
FADAt), x(t) = COUPLE) = ACEP)
= m(er(x(e), PAE) (20)
Clearly, ifthe unit is down (ie. x(¢) <0), then
GAPdt), x(t) = 0 (21)
If the unit is up, this non-start-up cost function does not
depend on the state x/(t) since the generation cost
C(pi(t)) and spinning reserve r,(x,(t), p,(E)) generally
depend only on the generation level if unit i is up. This